Export-banned Nvidia chips are behind China’s most powerful AI models — uh-oh, Jensen!

Unemployment Labs, a nonprofit research group led by former OpenAI researcher Will DePue, has published an analysis claiming that the most closely watched open-weight models from Chinese labs were trained on Nvidia Blackwell GPUs, which US export controls bar from shipment to China or Chinese-owned firms. The models examined include releases from DeepSeek, Moonshot AI, MiniMax and Ant Group. The finding, reported by TechCrunch, lands amid a widening scandal around Nvidia chief executive Jensen Huang, who left the company this week after admitting to an affair with Alice Huang, an executive at Megaspeed, a company implicated in smuggling Nvidia chips into China.
The technical method is the interesting part for developers. DePue did not rely on supply-chain records or whistleblowers; he analyzed the published weights of the open models themselves. Weights are the trained parameters that define how a model behaves. The claim is that training leaves detectable traces in those weights that reflect the configuration of the hardware used, and that the fingerprints in models such as Moonshot's Kimi K3 releases and DeepSeek's V4 point to Blackwell. This kind of weight forensics is inherently inferential: a fingerprint indicates the hardware was used at some stage, not that the model was trained exclusively on it, and the reporting acknowledges as much. Treat the result as strong circumstantial evidence rather than proof of a specific training run.
Even accepting the fingerprints, their meaning is contested. The chips could have sat in an offshore data center rather than on mainland Chinese soil, and the reporting notes that this route has been a popular way for Chinese firms to access compute, driving a buildout in Malaysia. Epoch AI, a nonprofit research organization, estimates roughly two million advanced chips entered China across 2024 and 2025 through schemes designed to obscure their final destination, and separately estimates 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025. Because the stated purpose of the bans is to stop Chinese labs training on the most powerful hardware, the distinction between ownership, location and end user is exactly where enforcement is weakest.
Sam Bresnick, a researcher at Georgetown's Center for Security and Emerging Technology, made that point to TechCrunch, arguing it makes little sense to bar Chinese companies from chips in mainland China, Hong Kong or Macau while permitting the same work in Malaysia. The policy backdrop has also shifted: Blackwell is banned outright for China and Chinese-owned firms, while the Trump administration allows export of the less powerful H200 provided shipments take a circuitous route that lets the US claim a 25% tariff. White House technology adviser Michael Kratsios said earlier this year, citing US intelligence, that Moonshot had used export-banned Nvidia chips.
The corporate story compounds the research. Jensen Huang's ties to Megaspeed, and Nvidia's own vouching for Megaspeed data centers after officials visited them, create exposure if investigators confirm smuggling occurred and that he knew of it. Megaspeed, led by Alice Huang, has pursued this trade for years via subsidiary Speedmatrix and has drawn investigations from the US and Singaporean governments; Alice Huang has also been alleged to be an agent of China's intelligence service. Bloomberg reported Friday that a company linked to her was involved in smuggling banned Nvidia chips to a data center in Shanghai. In May, short seller Culper published a report arguing Nvidia had a China problem because a meaningful share of revenue depended on Chinese demand through remote data centers or smuggling, and detailing Megaspeed's relationship with Nvidia. Culper's bet, notably, rested less on a US crackdown than on Beijing's push to build domestic chips, which could eventually erase billions in Chinese demand.
A further wrinkle: some 600,000 H200 chips are reportedly stockpiled in Hong Kong awaiting shipment into China, with no consensus on why they have not been cleared. Stanford research scholar Graham Webster said he doubts the holdup is meant to spur domestic production given the scale of demand; other theories include positioning for a trade deal involving large US imports, or concern about surveillance capability in the chips. Meanwhile, Commerce Secretary Howard Lutnick, whose agencies license chip exports, could reportedly lose his job over reports that investigators missed suspicious buyers and were too close to Nvidia.
Why it matters: For developers and product teams building on Chinese open-weight models, the practical question is reliability of supply rather than legality of the model itself—weight forensics like this can surface hardware provenance without any cooperation from the lab, which makes provenance claims testable in a way they were not before. If the technique holds up, expect model cards and procurement reviews to face pressure to document training hardware, and expect US enforcement to shift from tracking who owns chips toward who uses them, a change that could complicate the offshore compute arrangements several labs depend on. That enforcement shift is inference, not a stated policy; the near-term reality is that the most capable open models remain widely downloadable, and the dispute is over the machines behind them.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
New research shows the most advanced AI models from China were trained on banned Nvidia chips.